Migrate to SIE

If you’re already running embedding or decision inference somewhere else, these guides give you the smallest path to running it on SIE. Working before/after code per provider, plus a mapping from every provider concept to its SIE counterpart.

Pick your starting point:

OpenAI Drop-in via OpenAI-compatible endpoint, or the native SIE SDK. Eliminate per-token cost and rate limits.

OpenAI Decisions Keep client.decisions.create. Change the base URL and the model; predicate, choice and score questions carry over.

TypeSafe Keep the TypeSafe SDK and your System One questions. Point it at SIE and name an SIE decision model.

Cohere Replace embed-v3 and rerank-v3.5 with catalog embedding models and bge-reranker-v2-m3. Self-hosted, no rate limits.

TEI (HuggingFace) Replace N single-model TEI containers with one SIE cluster. Typed sparse and multivector outputs in one call.

Infinity Same OpenAI-compatible API, multi-model in one process, with managed deployment tooling.

Fastembed Move from in-process ONNX to out-of-process serving. Share GPU memory across app processes.

Modal Consolidate N Modal @app.function endpoints into one SIE deployment. Flat cost, no cold starts.

How to verify a migration

Each guide ships before/after code in the page. Run both legs on a small sample from your own domain and check the output looks sane: embeddings for an embedding migration, answers for a decision migration. That is a sanity check, not sign-off.

For a decision migration (OpenAI Decisions, TypeSafe), sign off on a labeled set: a few hundred inputs from your own traffic with the answer you expect to each question. Run both legs and compare accuracy per question. For a predicate, also pick the probability threshold you act on from that set, because each model’s probabilities fall in their own range.

For an embedding migration, run your own retrieval eval against both legs:

• Same checkpoint (Fastembed, TEI, Infinity, Modal-with-same-model). Cosine should sit at 0.999 or higher across items. If it doesn’t, the config differs (pooling, normalization, dtype); the guide’s caveats section calls out where.
• Different model (OpenAI → E5, Cohere → Stella or E5). Absolute cosine carries no signal across spaces. Run recall@k on a labeled set you trust, or a BEIR/MTEB slice that resembles your domain.

Choosing a target model

Source model	Closest SIE model	Re-embed
text-embedding-3-small	intfloat/e5-base-v2	yes
text-embedding-3-large	Alibaba-NLP/gte-Qwen2-1.5B-instruct	yes
embed-english-v3.0 (Cohere)	NovaSearch/stella_en_400M_v5	yes
rerank-v3.0 (Cohere)	BAAI/bge-reranker-v2-m3	n/a
TEI / Infinity / Fastembed bge-*	same checkpoint on SIE	no
sentence-transformers on Modal	same checkpoint on SIE	no

Browse the full model catalog for everything SIE serves out of the box.

Need a guide that isn’t here?

Open an issue at superlinked/sie.